feat(kernel): surface kernel logs through Python logging on use_kernel

[vikrantpuppala]

Summary

On the use_kernel=True path, the Rust kernel's logs now surface through Python's stdlib logging automatically — no extra setup. Kernel logs land under the databricks.sql.kernel logger, a child of the connector's databricks.sql.* namespace, so a customer's existing logging config just works:

importlogginglogging.getLogger("databricks.sql").setLevel(logging.DEBUG) # connector + kernel

Before this, kernel-backed sessions emitted nothing into Python logging — the kernel produced tracing events but nothing on the Python path consumed them.

Changes

• backend/kernel/_errors.py calls databricks_sql_kernel.init_logging() once when the kernel extension loads (the canonical kernel-import site). The call is getattr-guarded, so an older kernel wheel without the function still works — just without kernel logs.
• tests/e2e/test_kernel_backend.py adds two creds-gated e2e tests (matching the existing kernel e2e convention):
  • test_kernel_logs_reach_python_logging — a query at DEBUG produces records on databricks.sql.kernel, and the pyo3 boundary surfaces under databricks.sql.kernel.pyo3.
  • test_kernel_log_level_is_respected — at WARNING, the chatty DEBUG kernel records are filtered out (proving level control, not unconditional forwarding).

Cross-layer visibility

With this, a single query shows the full vertical in one stream:

databricks.sql.session DEBUG Creating kernel-backed client for use_kernel=True
databricks.sql.kernel.pyo3 DEBUG Session.__new__: opening kernel session (GIL released)
databricks.sql.kernel DEBUG CreateSession: POST .../sessions (attempt 1/6)
databricks.sql.kernel INFO kernel.session; session_id=...
databricks.sql.client DEBUG Cursor.execute(operation=SELECT 1 AS a)
databricks.sql.kernel.pyo3 DEBUG Statement.execute: running statement (GIL released)
databricks.sql.kernel DEBUG ExecuteStatement: POST .../statements (attempt 1/6)

Dependency

Requires the companion kernel/pyo3 change that exposes init_logging(): databricks/databricks-sql-kernel#120. Until a kernel wheel containing that function is installed, this is a no-op (the getattr guard), so it's safe to merge independently.

Testing

Both e2e tests pass against a live warehouse with a locally-built kernel wheel (maturin develop).

🤖 Generated with Claude Code

[gopalldb]

🟠 P1 (Important — should fix before merge)

1. init_logging() call is not wrapped against exceptions/panics — _errors.py:74-76 (confirmed: bare _kernel_init_logging()). The getattr guards a missing function, not a throwing one, and the surrounding
   try/except ImportError only catches the import. A pyo3 panic surfaces as pyo3_runtime.PanicException — a BaseException, not caught by except Exception — and would propagate out of module import, failing
   every use_kernel=True connection in the process over a non-essential logging feature. → wrap in try/except BaseException (deliberate, since pyo3 panics derive from it), log a debug breadcrumb, and continue
   without the bridge.
2. The "idempotent on the Rust side" claim is load-bearing and unverifiable here — _errors.py:71-72. Safety rests on the Rust side using try_init/a guard rather than
   pyo3_log::init()/tracing_subscriber...init(), both of which panic if a global logger was already set. The logger is process-global, so if the host app (or another lib) already installed a log/tracing
   subscriber, a non-idempotent init panics → lands on P1#1. → land the try/except (neutralizes it regardless of the Rust impl), and confirm/tighten the comment to the exact contract ("safe to call repeatedly;
   never panics").
3. Forwarding kernel DEBUG logs widens the data-exposure surface — once installed, any app setting databricks.sql to DEBUG (a common troubleshooting step) receives whatever the Rust kernel emits at
   DEBUG/TRACE — commonly SQL text, bound parameter values, request URLs, and Authorization headers. None of that is visible/redactable from this repo, so the connector can't assert no leakage. Unlike
   pure-Python paths where the connector controls/redacts its own log content, this routes kernel internals into app logs/aggregation. → audit what the kernel emits at DEBUG/TRACE for the databricks.sql.kernel
   target; confirm credentials/auth headers are never logged; decide whether SQL/param exposure at DEBUG is acceptable, and if not, cap the forwarded level (e.g. INFO+) until the Rust side guarantees redaction.
   Document the exposure in user docs.

🟡 P2 (Minor)

• Level-filtering test proves caplog handler filtering, not source-side suppression — test_kernel_backend.py:173-194. caplog.at_level(WARNING, ...) sets the logger level and attaches a handler, so a DEBUG
  record that crossed the FFI would still be dropped by Python before reaching caplog.records. The test can't distinguish "kernel suppressed at source" from "Python filtered after FFI" — false confidence that
  a chatty TRACE loop avoids per-record FFI + LogRecord cost. → assert source-side suppression explicitly (emit counter) or soften the docstring + note the per-record FFI cost.
• Hard-asserting databricks.sql.kernel.pyo3 records — :165-167 couples the connector test suite to the kernel's internal pyo3-log target naming; a benign kernel change breaks e2e. → treat .pyo3 as a
  soft/optional check.
• Skip hygiene — verified clean: importorskip + fake-stub file guard + kernel_conn_params gating mean both new tests skip cleanly without creds/wheel.

[vikrantpuppala]

Thanks @gopalldb — these are great catches. Dispositions below; pushed in ca9234b.

P1#1 — bare init_logging() vulnerable to a throwing/panicking call ✅ Fixed

You're right. Wrapped in try/except BaseException (deliberately broad, since a PyO3 panic surfaces as pyo3_runtime.PanicException, a BaseException), logging a debug breadcrumb and continuing without the bridge. Logging is non-essential and must never take down a use_kernel=True connection.

P1#2 — idempotency claim load-bearing & unverifiable here ✅ Covered

The try/except above neutralizes this regardless of the Rust impl, as you suggested. FWIW the kernel side is safe — it guards with a OnceLock and uses pyo3_log::Logger::install() which returns a Result (matched, returns false on Err), so it returns rather than panics on a double install — but we no longer rely on that. Tightened the comment to state the guard holds regardless.

P1#3 — DEBUG data-exposure surface ✅ Audited; documented; no level cap

I audited what the kernel emits for the databricks::sql::kernel target:

• Auth headers / tokens: never logged. No Authorization/Bearer/token-value logging. OAuth lines log cache state ("loaded cached token with refresh_token"), not secrets.
• Bound parameter values: never logged.
• SQL text: logged at DEBUG (ExecuteStatement + metadata queries).
• URLs: logged, but they're API endpoints (/api/2.0/sql/statements), not credential-bearing.

On SQL-at-DEBUG: this is parity with the existing connector, which already logs both SQL and parameter values at DEBUG (client.py:1331: Cursor.execute(operation=%s, parameters=%s)). The kernel path is actually more conservative — SQL but not params. So this isn't a new exposure class for the driver; capping the kernel to INFO+ would (a) break that parity, (b) gut the feature's primary use case (debugging queries), and (c) still leave the connector's own DEBUG SQL/param logging in place. I'd rather keep parity and document it. Happy to add a note to user-facing docs that databricks.sql.kernel at DEBUG includes SQL text — let me know if you'd like that in this PR or a docs follow-up.

P2 — level test proves handler filtering, not source suppression ✅ Docstring softened

Correct. Reworded the docstring to scope the claim to the effective customer-visible outcome (sub-threshold records don't surface), and noted that source-side suppression / per-record FFI cost avoidance is a kernel-side concern, not asserted here.

P2 — hard-asserting databricks.sql.kernel.pyo3 couples to kernel internals ✅ Softened

Downgraded to a warnings.warn so a benign kernel change to the boundary breadcrumb target doesn't break the connector e2e. The core databricks.sql.kernel contract stays hard-asserted.

P2 — skip hygiene

Thanks for confirming.